Opus 4.8
Anthropic发布的Claude Opus系列模型，在OSWorld 2.0基准中被GPT-5.6 Sol超越
Core Facts
GLM-5.3 Flash's capabilities are on par with Anthropic's Opus 4.8, but its inference cost is only a fraction of the latter's.
DeepSeek has launched the experimental multimodal model V4-Flash-Vision-EXP on its API platform, with text capabilities on par with V4 Flash and multimodal agent benchmark performance close to Opus 4.8 levels.
Timeline (last 90 days)
Opus 5.5's safety mechanism interrupted the session while analyzing the app's permission mappings and proposed switching to Opus 4.8.
Versions Opus 4.7, 4.8, and 5.0 exhibited a decline in conversational quality, a heavier AI-like feel, and deteriorating task execution performance.
Existing prompts written for Opus 5 run on Opus 5.5 without modification.
An overseas developer blogger who has long championed Claude Code believes that the actual performance of DeepSeek's Pro version is comparable to Anthropic's Opus 4.8.
The blogger believes that DeepSeek V4 Pro's performance may have already surpassed Opus 4.8.
An overseas tech blogger believes that the latest DeepSeek is already on par with Anthropic's Opus 4.8 on programming tasks, but at only a fraction of the cost.
The blogger believes the DeepSeek V4 Pro version released in the past two weeks may already be superior to Opus 4.8, but admits to being uncertain about specific benchmark data.
Under a 15-minute time limit per problem, the proof agent achieved a successful proof rate of 34.9% (157/450) on StochBench.
The research team systematically tested StochBench using an agent based on Opus 4.8.
In tests with the strategy game Slay the Spire, the performance improvement brought by persistent file memory was three times that of Opus 4.8.
9 more timeline events
All Facts (12)
GLM-5.3 Flash's capabilities are on par with Anthropic's Opus 4.8, but its inference cost is only a fraction of the latter's.
75%VerifiedDeepSeek has launched the experimental multimodal model V4-Flash-Vision-EXP on its API platform, with text capabilities on par with V4 Flash and multimodal agent benchmark performance close to Opus 4.8 levels.
85%UnverifiedClaude Opus 5 scored 30.2% on ARC-AGI-3, compared to only 1.5% for the previous generation.
60%UnverifiedVersions Opus 4.7, 4.8, and 5.0 exhibited a decline in conversational quality, a heavier AI-like feel, and deteriorating task execution performance.
50%UnverifiedExisting prompts written for Opus 5 run on Opus 5.5 without modification.
50%UnverifiedThe research team systematically tested StochBench using an agent based on Opus 4.8.
50%UnverifiedUnder a 15-minute time limit per problem, the proof agent achieved a successful proof rate of 34.9% (157/450) on StochBench.
50%UnverifiedClaude Fable 5 was released only 11 days after the previous generation Opus 4.8.
50%UnverifiedIn tests with the strategy game Slay the Spire, the performance improvement brought by persistent file memory was three times that of Opus 4.8.
50%Unverified论文在三大厂商验证有效：Anthropic用Haiku 4.5解码Opus 4.8，OpenAI用GPT-5 Mini解码GPT-5，Google用Gemini解码其Flash版本，且解码后推理链token数量与API报告完全吻合
50%UnverifiedOpus 5.5's safety mechanism interrupted the session while analyzing the app's permission mappings and proposed switching to Opus 4.8.
50%UnverifiedClaude Code启动时使用Fable 5模型，但触发安全检查后回退到Opus 4.8作为主编排器，再由它调度Fable 5智能体执行任务
50%